[PATCH v4 0/7] sched/fair: improve scan efficiency of SIS

From: Abel Wu
Date: Sun Jun 19 2022 - 08:05:18 EST


The wakeup fastpath (select_idle_sibling or SIS) plays an important role
in maximizing the usage of cpu resources and can greatly affect overall
performance of the system.

The SIS tries to find an idle cpu inside that LLC to place the woken-up
task. The cache hot cpus will be checked first, then other cpus of that
LLC (domain scan) if the hot ones are not idle.

The domain scan works well under light workload by simply traversing
the cpus of the LLC due to lots of idle cpus can be available. But this
doesn’t scale well once the LLC gets bigger and the load increases, so
SIS_PROP was born to limit the scan cost. For now SIS_PROP just limits
the number of cpus to be scanned, but the way of how it scans is not
changed.

This patchset introduces the SIS filter to help improving scan efficiency
when scan depth is limited. The filter only contains the unoccupied cpus,
and is updated during SMT level load balancing. It is expected that the
more overloaded the system is, the less cpus will be scanned.

SIS_PROP
=======

The filter should be put under SIS_PROP which is not now, because this
sched-feature is under rework or been replaced (by SIS_UTIL [1]). Will
be adjusted once that work is ready.

[1] https://lore.kernel.org/all/20220612163428.849378-1-yu.c.chen@xxxxxxxxx/

Entrance: Load balance vs. Tick
===============================

Now we update the filter during load balancing at SMT level, while it was
done at the tick (patch v1-v3). There are several reasons:

- The periodic load balancing has the admission control that only the
first idle cpu (if any) or the balance cpu can trigger balance, so
the complexity is reduced by O(n) in the domain scope. While on the
other hand, the complexity of per-cpu updating at the tick is not
reduced, and can cause heavy traffic on the LLC cache.

- In order to keep the filter relatively fresh, the cpu idle path needs
to be taken into consideration as well. Since load_balance is the
common path for the periodic and newly-idle balance, we can simplify
and concentrate on load_balance itself. While the tick way both tick
and idle path need to be modified.

- The calculation of SMT state can re-use the intermediate results of
the load balance, so it’s cheap without sacrifice.

But the tick way also has its advantages in the idle path that the cpus
set to filter are guaranteed to be idle compared to load_balance in which
the filter is updated before task migrations (so the idle dst_cpu can be
fed with tasks). This can be fixed by false positive correction at later
discussion.

Filter: Busy vs. Overloaded
===========================

In preious patch versions, rather than the busy or unoccupied cpus, only
the overloaded ones will be set into the filter because of its stableness.
For example, the state of the unoccupied cpus (non-idle tasks == 0) or the
busy cpus (== 1) can be easily changed by a short running workload. And
besides, wakup won’t pollute the filter since add tasks to the already
overloaded cpus can’t change their states.

But it’s different in this patchset in which the load balance cpu will be
responsible for updating its core state. The state will be set to has_icpus
if unoccupied cpus exist no matter what the other cpu states are. So the
overloaded state of a core is also unstable. Given this, the semantics of
busy is much preferred than overloaded due to its accuracy.

The only problem comes to cache, again. To avoid frequent update thanks
to the variance, the false positive bits in the filter are allowed to exist
for a while, which will be talked about later.

Incremental Propagation
=======================

By design, the unoccupied cpus will be set (or propagate) to the filter.
But it’s heavy to update the whole SMT state every time. What we actually
do is only to propagate one unoccupied cpu of this core to the filter at
a time, so called incremental propagation. It’s not just a tradeoff, it
also helps spreading load to different cores.

The cpu to be propagated needs to be selected carefully. If the load
balance cpu is chosen, the filter might be polluted because that cpu
could be fed with tasks soon. But it’s not wise either to wait to the
end of load balancing as we need to do it as early as possible to keep
the filter fresh. The solution is false positive correction too.

Accuracy: False Positive Correction
===================================

The filter is supposed to contain unoccupied cpus. There are two main
cases that will downgrade the filter:

- False negtive: there are unoccupied cpus not represented in the
filter. This is the case we should try to eliminate because these
cpus will be out of reach in the wakeup leading to in-sufficient
use of cpu resources. So in practical the unoccupied cpus will be
aggressively propagated to the filter.

- False positive: some of the cpus in the filter are not unoccupied
any more. We allow these cpus to stay in the filter for some time
to relieve cache pressure. It's a tradeoff.

A correction is needed when the false positive case is continuously
detected. See patch 7 for detailed info.

SMT 2/4/8
=========

The incremental propagation is relatively friendly to SMT2, but not for
the others. As the core grows bigger, the more false negtive cpus could
exist. It would be appreciated if anyone can help testing on SMT 4/8
machines.

Benchmark
=========

Tests are done in an Intel Xeon(R) Platinum 8260 CPU@2.40GHz machine
with 2 NUMA nodes each of which has 24 cores with SMT2 enabled, so 96
CPUs in total.

All of the benchmarks are done inside a normal cpu cgroup in a clean
environment with cpu turbo disabled.

Based on tip sched/core f3dd3f674555 (v5.19-rc2).

Results
=======

hackbench-process-pipes
vanilla filter
Amean 1 0.2357 ( 0.00%) 0.2540 ( -7.78%)
Amean 4 0.6403 ( 0.00%) 0.6363 ( 0.62%)
Amean 7 0.7583 ( 0.00%) 0.7367 ( 2.86%)
Amean 12 1.0703 ( 0.00%) 1.0520 ( 1.71%)
Amean 21 2.0363 ( 0.00%) 1.9610 * 3.70%*
Amean 30 3.2487 ( 0.00%) 2.9083 * 10.48%*
Amean 48 6.3620 ( 0.00%) 4.8543 * 23.70%*
Amean 79 8.3653 ( 0.00%) 7.1077 * 15.03%*
Amean 110 9.8370 ( 0.00%) 8.5740 * 12.84%*
Amean 141 11.4667 ( 0.00%) 10.8750 * 5.16%*
Amean 172 13.4433 ( 0.00%) 12.6443 * 5.94%*
Amean 203 15.8970 ( 0.00%) 14.9143 * 6.18%*
Amean 234 17.9643 ( 0.00%) 16.9123 * 5.86%*
Amean 265 20.3910 ( 0.00%) 19.2040 * 5.82%*
Amean 296 22.5253 ( 0.00%) 21.2547 * 5.64%*

hackbench-process-sockets
vanilla filter
Amean 1 0.4177 ( 0.00%) 0.4133 ( 1.04%)
Amean 4 1.4397 ( 0.00%) 1.4240 * 1.09%*
Amean 7 2.4720 ( 0.00%) 2.4310 * 1.66%*
Amean 12 4.1407 ( 0.00%) 4.0683 * 1.75%*
Amean 21 7.0550 ( 0.00%) 6.8830 * 2.44%*
Amean 30 9.9633 ( 0.00%) 9.7750 * 1.89%*
Amean 48 15.9837 ( 0.00%) 15.5313 * 2.83%*
Amean 79 26.7740 ( 0.00%) 26.2703 * 1.88%*
Amean 110 37.2913 ( 0.00%) 36.5433 * 2.01%*
Amean 141 47.8937 ( 0.00%) 46.5300 * 2.85%*
Amean 172 58.0273 ( 0.00%) 56.4530 * 2.71%*
Amean 203 68.2530 ( 0.00%) 66.3320 * 2.81%*
Amean 234 78.8987 ( 0.00%) 76.8497 * 2.60%*
Amean 265 89.1520 ( 0.00%) 86.8213 * 2.61%*
Amean 296 99.6920 ( 0.00%) 96.9707 * 2.73%*

hackbench-thread-pipes
vanilla filter
Amean 1 0.2647 ( 0.00%) 0.2633 ( 0.50%)
Amean 4 0.6290 ( 0.00%) 0.6607 ( -5.03%)
Amean 7 0.7850 ( 0.00%) 0.7870 ( -0.25%)
Amean 12 1.3347 ( 0.00%) 1.2577 ( 5.77%)
Amean 21 3.1233 ( 0.00%) 2.4613 * 21.20%*
Amean 30 5.7120 ( 0.00%) 3.6847 * 35.49%*
Amean 48 8.1947 ( 0.00%) 6.2670 * 23.52%*
Amean 79 9.1750 ( 0.00%) 8.0640 * 12.11%*
Amean 110 10.6300 ( 0.00%) 9.5583 * 10.08%*
Amean 141 12.7490 ( 0.00%) 12.0167 * 5.74%*
Amean 172 15.1567 ( 0.00%) 14.1570 * 6.60%*
Amean 203 17.5160 ( 0.00%) 16.7883 ( 4.15%)
Amean 234 19.8710 ( 0.00%) 19.5370 ( 1.68%)
Amean 265 23.2700 ( 0.00%) 21.4017 * 8.03%*
Amean 296 25.4093 ( 0.00%) 23.9943 * 5.57%*

hackbench-thread-sockets
vanilla filter
Amean 1 0.4467 ( 0.00%) 0.4347 ( 2.69%)
Amean 4 1.4757 ( 0.00%) 1.4533 * 1.51%*
Amean 7 2.5320 ( 0.00%) 2.4993 * 1.29%*
Amean 12 4.2617 ( 0.00%) 4.1780 * 1.96%*
Amean 21 7.2397 ( 0.00%) 7.0660 * 2.40%*
Amean 30 10.2200 ( 0.00%) 9.9810 * 2.34%*
Amean 48 16.2623 ( 0.00%) 16.0483 * 1.32%*
Amean 79 27.4307 ( 0.00%) 26.8410 * 2.15%*
Amean 110 37.8993 ( 0.00%) 37.3223 * 1.52%*
Amean 141 48.3890 ( 0.00%) 47.4823 * 1.87%*
Amean 172 58.9887 ( 0.00%) 57.7753 * 2.06%*
Amean 203 69.5853 ( 0.00%) 68.0563 * 2.20%*
Amean 234 80.0743 ( 0.00%) 78.4857 * 1.98%*
Amean 265 90.5473 ( 0.00%) 89.3363 * 1.34%*
Amean 296 101.3857 ( 0.00%) 99.7717 * 1.59%*

schbench
vanilla filter
Lat 50.0th-qrtle-1 6.00 ( 0.00%) 6.00 ( 0.00%)
Lat 75.0th-qrtle-1 8.00 ( 0.00%) 8.00 ( 0.00%)
Lat 90.0th-qrtle-1 9.00 ( 0.00%) 8.00 ( 11.11%)
Lat 95.0th-qrtle-1 9.00 ( 0.00%) 8.00 ( 11.11%)
Lat 99.0th-qrtle-1 10.00 ( 0.00%) 9.00 ( 10.00%)
Lat 99.5th-qrtle-1 11.00 ( 0.00%) 9.00 ( 18.18%)
Lat 99.9th-qrtle-1 11.00 ( 0.00%) 9.00 ( 18.18%)
Lat 50.0th-qrtle-2 6.00 ( 0.00%) 7.00 ( -16.67%)
Lat 75.0th-qrtle-2 7.00 ( 0.00%) 8.00 ( -14.29%)
Lat 90.0th-qrtle-2 8.00 ( 0.00%) 9.00 ( -12.50%)
Lat 95.0th-qrtle-2 8.00 ( 0.00%) 10.00 ( -25.00%)
Lat 99.0th-qrtle-2 9.00 ( 0.00%) 11.00 ( -22.22%)
Lat 99.5th-qrtle-2 9.00 ( 0.00%) 11.00 ( -22.22%)
Lat 99.9th-qrtle-2 9.00 ( 0.00%) 11.00 ( -22.22%)
Lat 50.0th-qrtle-4 9.00 ( 0.00%) 8.00 ( 11.11%)
Lat 75.0th-qrtle-4 10.00 ( 0.00%) 10.00 ( 0.00%)
Lat 90.0th-qrtle-4 11.00 ( 0.00%) 11.00 ( 0.00%)
Lat 95.0th-qrtle-4 12.00 ( 0.00%) 11.00 ( 8.33%)
Lat 99.0th-qrtle-4 13.00 ( 0.00%) 13.00 ( 0.00%)
Lat 99.5th-qrtle-4 13.00 ( 0.00%) 16.00 ( -23.08%)
Lat 99.9th-qrtle-4 13.00 ( 0.00%) 19.00 ( -46.15%)
Lat 50.0th-qrtle-8 11.00 ( 0.00%) 11.00 ( 0.00%)
Lat 75.0th-qrtle-8 14.00 ( 0.00%) 14.00 ( 0.00%)
Lat 90.0th-qrtle-8 16.00 ( 0.00%) 16.00 ( 0.00%)
Lat 95.0th-qrtle-8 17.00 ( 0.00%) 17.00 ( 0.00%)
Lat 99.0th-qrtle-8 22.00 ( 0.00%) 19.00 ( 13.64%)
Lat 99.5th-qrtle-8 28.00 ( 0.00%) 23.00 ( 17.86%)
Lat 99.9th-qrtle-8 31.00 ( 0.00%) 42.00 ( -35.48%)
Lat 50.0th-qrtle-16 17.00 ( 0.00%) 17.00 ( 0.00%)
Lat 75.0th-qrtle-16 23.00 ( 0.00%) 23.00 ( 0.00%)
Lat 90.0th-qrtle-16 26.00 ( 0.00%) 27.00 ( -3.85%)
Lat 95.0th-qrtle-16 28.00 ( 0.00%) 29.00 ( -3.57%)
Lat 99.0th-qrtle-16 32.00 ( 0.00%) 33.00 ( -3.12%)
Lat 99.5th-qrtle-16 37.00 ( 0.00%) 35.00 ( 5.41%)
Lat 99.9th-qrtle-16 54.00 ( 0.00%) 46.00 ( 14.81%)
Lat 50.0th-qrtle-32 30.00 ( 0.00%) 29.00 ( 3.33%)
Lat 75.0th-qrtle-32 43.00 ( 0.00%) 42.00 ( 2.33%)
Lat 90.0th-qrtle-32 51.00 ( 0.00%) 49.00 ( 3.92%)
Lat 95.0th-qrtle-32 54.00 ( 0.00%) 51.00 ( 5.56%)
Lat 99.0th-qrtle-32 61.00 ( 0.00%) 57.00 ( 6.56%)
Lat 99.5th-qrtle-32 64.00 ( 0.00%) 60.00 ( 6.25%)
Lat 99.9th-qrtle-32 72.00 ( 0.00%) 82.00 ( -13.89%)
Lat 50.0th-qrtle-47 44.00 ( 0.00%) 45.00 ( -2.27%)
Lat 75.0th-qrtle-47 64.00 ( 0.00%) 65.00 ( -1.56%)
Lat 90.0th-qrtle-47 75.00 ( 0.00%) 77.00 ( -2.67%)
Lat 95.0th-qrtle-47 81.00 ( 0.00%) 82.00 ( -1.23%)
Lat 99.0th-qrtle-47 92.00 ( 0.00%) 98.00 ( -6.52%)
Lat 99.5th-qrtle-47 101.00 ( 0.00%) 114.00 ( -12.87%)
Lat 99.9th-qrtle-47 271.00 ( 0.00%) 167.00 ( 38.38%)

netperf-udp
vanilla filter
Hmean send-64 199.12 ( 0.00%) 201.32 ( 1.11%)
Hmean send-128 396.22 ( 0.00%) 397.01 ( 0.20%)
Hmean send-256 777.80 ( 0.00%) 783.96 ( 0.79%)
Hmean send-1024 2972.62 ( 0.00%) 3011.87 * 1.32%*
Hmean send-2048 5600.64 ( 0.00%) 5730.50 * 2.32%*
Hmean send-3312 8757.45 ( 0.00%) 8703.62 ( -0.61%)
Hmean send-4096 10578.90 ( 0.00%) 10590.93 ( 0.11%)
Hmean send-8192 17051.22 ( 0.00%) 17189.62 * 0.81%*
Hmean send-16384 27915.16 ( 0.00%) 27816.01 ( -0.36%)
Hmean recv-64 199.12 ( 0.00%) 201.32 ( 1.11%)
Hmean recv-128 396.22 ( 0.00%) 397.01 ( 0.20%)
Hmean recv-256 777.80 ( 0.00%) 783.96 ( 0.79%)
Hmean recv-1024 2972.62 ( 0.00%) 3011.87 * 1.32%*
Hmean recv-2048 5600.64 ( 0.00%) 5730.49 * 2.32%*
Hmean recv-3312 8757.45 ( 0.00%) 8703.61 ( -0.61%)
Hmean recv-4096 10578.90 ( 0.00%) 10590.93 ( 0.11%)
Hmean recv-8192 17051.21 ( 0.00%) 17189.57 * 0.81%*
Hmean recv-16384 27915.08 ( 0.00%) 27815.86 ( -0.36%)

netperf-tcp
vanilla filter
Hmean 64 811.07 ( 0.00%) 835.46 * 3.01%*
Hmean 128 1614.86 ( 0.00%) 1652.27 * 2.32%*
Hmean 256 3131.16 ( 0.00%) 3119.01 ( -0.39%)
Hmean 1024 10286.12 ( 0.00%) 10333.64 ( 0.46%)
Hmean 2048 16231.88 ( 0.00%) 17141.88 * 5.61%*
Hmean 3312 20705.91 ( 0.00%) 21703.49 * 4.82%*
Hmean 4096 22650.75 ( 0.00%) 23904.09 * 5.53%*
Hmean 8192 27984.06 ( 0.00%) 29170.57 * 4.24%*
Hmean 16384 32816.85 ( 0.00%) 33351.41 * 1.63%*

tbench4 Throughput
vanilla filter
Hmean 1 300.07 ( 0.00%) 302.52 * 0.82%*
Hmean 2 617.72 ( 0.00%) 598.45 * -3.12%*
Hmean 4 1213.99 ( 0.00%) 1206.36 * -0.63%*
Hmean 8 2373.78 ( 0.00%) 2372.28 * -0.06%*
Hmean 16 4777.82 ( 0.00%) 4711.44 * -1.39%*
Hmean 32 7182.50 ( 0.00%) 7718.15 * 7.46%*
Hmean 64 8611.44 ( 0.00%) 9409.29 * 9.27%*
Hmean 128 18102.63 ( 0.00%) 20650.23 * 14.07%*
Hmean 256 18029.28 ( 0.00%) 20611.03 * 14.32%*
Hmean 384 17986.44 ( 0.00%) 19361.29 * 7.64%*

Conclusion
==========

There seems like not much obvious regressions except the hackbench pipe
tests with minor groups like 1/4, and some of the benchmarks showed great
improvement.

As preious said, the domain scan works well under light workload by simply
traversing the cpus of the LLC due to lots of idle cpus can be available.
So in this case, the cost of maintaining the filter can hardly contribute
to the scan efficiency, but rather adds to the scheduling overhead slowing
down the system.

One reasonable way to fix it is to disable the filter when the system is
not loaded. Three possible solutions come to my mind:

a. Disable the filter when the number of unoccupied cpus in the LLC
is more than a threshold. This is straight but not easy to define a
proper threshold for different topologies or workloads. So I gave up
on this quickly.

b. Enable the filter when average idle time reduces. This is what the
SIS_PROP relies on. I did a test on enabling the filter when nr=4,
but the result is not convincing..

c. Enable the filter when scan efficiency drops to a threshold. Like a,
the threshold is hard to define.

I am still working on this and open to discussion and suggestions.
Comments and tests are appreciated!

v3 -> v4:
- update filter in load_balance rather than in the tick
- now the filter contains unoccupied cpus rather than overloaded ones
- added mechanisms to deal with the false positive cases

v2 -> v3:
- removed sched-idle balance feature and focus on SIS
- take non-CFS tasks into consideration
- several fixes/improvement suggested by Josh Don

v1 -> v2:
- several optimizations on sched-idle balancing
- ignore asym topos in can_migrate_task
- add more benchmarks including SIS efficiency
- re-organize patch as suggested by Mel

v3: https://lore.kernel.org/lkml/53fcde27-7dd2-5150-633b-4e2acc37bb6f@xxxxxxxxxxxxx/
v2: https://lore.kernel.org/lkml/20220409135104.3733193-1-wuyun.abel@xxxxxxxxxxxxx/
v1: https://lore.kernel.org/lkml/20220217154403.6497-5-wuyun.abel@xxxxxxxxxxxxx/

Abel Wu (7):
sched/fair: default to false in test_idle_cores
sched/fair: remove redundant check in select_idle_smt
sched/fair: avoid double search on same cpu
sched/fair: remove useless check in select_idle_core
sched/fair: skip SIS domain search if fully busy
sched/fair: skip busy cores in SIS search
sched/fair: de-entropy for SIS filter

include/linux/sched/topology.h | 62 ++++++++-
kernel/sched/fair.c | 233 +++++++++++++++++++++++++++++----
kernel/sched/topology.c | 12 +-
3 files changed, 277 insertions(+), 30 deletions(-)

--
2.31.1